Clarify specification around type conversion via Column.get(int, Class)

[lukaseder]

Feature Request

See also discussion here:
https://groups.google.com/g/r2dbc/c/0fpr5DY3big

Is your feature request related to a problem? Please describe

In https://r2dbc.io/spec/0.9.0.M1/spec/html/#rows, the specification claims that

When you specify a target type, the R2DBC driver tries to convert the value to the target type

This specification is a bit vague and allows for both assumptions:

• The behaviour is deliberately implementation-specific, and clients will have to implement conversions themselves, or
• The behaviour is implementation-specific to some extent, but some "reasonable" conversions can be expeted to work

Describe the solution you'd like

I'm assuming that some "reasonable" conversions can be safely expected to work on all drivers, but this is currently not the case, see e.g.:

• Cannot convert between numeric types when fetching a value mirromutth/r2dbc-mysql#177
• Integer literal cannot be converted to non Integer.class numeric types r2dbc-h2#190

I'd like the specification to be more specific, not having an opinion in which direction it should go, but as a client implementation, I would like to have certainty that certain conversions are can safely be expected to work now, and in the future.

[lukaseder]

I'd like to double down on this lack of specification clarity here. We only have a few drivers yet, and already I don't know if I should use byte[].class (the JDBC type), ByteBuffer.class (the R2dbcType type) or io.r2dbc.spi.Blob. For example:

create table t (b blob);
insert into t values (null);

And then:

Class<?> klass = ...?

System.out.println(
Flux.from(cf.create())
    .flatMap(c -> c.createStatement("select b from t").execute())
    .flatMap(it -> it.map((r, m) -> Optional.ofNullable(r.get(0, klass))))
    .collectList()
    .block()
);

What do I put for klass? The oracle driver (to my surprise) only supports io.r2dbc.spi.Blob. I mean, that's not an unreasonable type to support, but I would have really expected R2dbcType.BLOB.getJavaType() to work (which maps to ByteBuffer). However, I don't know if I should report a bug to https://github.com/oracle/oracle-r2dbc

[mp911de]

For real blobs, that can contain a significant amount of data (an amount that doesn't easily fit into your heap size), please use Blob, as per https://r2dbc.io/spec/0.8.5.RELEASE/spec/html/#datatypes.mapping.binary. ByteBuffer should be generally supported but there's a limitation for databases that use locators such as Oracle. A locator requires additional database interaction which may cause a blocking behavior upon Row.get(0, ByteBuffer.class). I think because of that database-specific limitation they decided to not support ByteBuffer to avoid surprises.

[lukaseder]

I don't remember having used those Oracle BLOB features, ever. When I stored images or PDFs in BLOBs, byte[] always sufficed with JDBC. Correctly streaming chunked blobs while implementing correct resource management is extremely hard, both for users and integrators. Maybe, a 5GB BLOB doesn't belong in the database but on a simple file system?

Anyway. Thanks for the clarification on this part.

[mp911de]

Maybe, a 5GB BLOB doesn't belong in the database but on a simple file system?

Fully agree. According to their database, a blob can hold "TB's of data" so mentioned that for full transparency and consideration.

[lukaseder]

I've raised the question in this discussion: oracle/oracle-r2dbc#17 (comment). Curious about the findings here.

[Michael-A-McMahon]

I'd really like for Oracle R2DBC to support the ByteBuffer mapping of BLOB, but so far I've been unable to identify a good way to do it. I posted my thoughts about this here:
oracle/oracle-r2dbc#32 (comment)

Maybe someone can think of a better solution? If so, then please post on the issue I've linked to above. Any discussion is welcome!

[Michael-A-McMahon]

Going back to the original topic, Section 14 of the R2DBC Specification provides a "guideline" of SQL to Java type mappings. If we can expect all R2DBC drivers to support these mappings, does this offer enough clarify an R2DBC consumer?